This is the mail archive of the
cgen@sourceware.org
mailing list for the CGEN project.
Re: Help needed for porting opcodes for CISC architecture
- From: Usha Gupta <gusha2960 at gmail dot com>
- To: "Frank Ch. Eigler" <fche at redhat dot com>
- Cc: cgen <cgen at sourceware dot org>
- Date: Mon, 17 Feb 2014 14:41:30 +0530
- Subject: Re: Help needed for porting opcodes for CISC architecture
- Authentication-results: sourceware.org; auth=none
- References: <CANcummWUhCCPZsVRR9EaeZ7Y4o0kK1E4bv9r48qBvPUych2yjA at mail dot gmail dot com> <y0mlhxipscw dot fsf at fche dot csb>
Thanks Frank.
I have got instructions upto 4 bytes being correctly assembled.
I was initially using simplification macro "dnf " for defining
instruction fields which does not allows specifying word-length and
word-offset.
The problem is resolved by specifying instruction fields using
define-ifield where I could define the total length of the instruction
where this specific instruction field is used .
For 5-byte insns, since opcode doesn't fit in CGEN_INSN_INT which is
defined in include/opcode/cgen.h as unsigned int , I changed it to
"unsigned long long". Is there a cleaner way to achieve this?
The instructions seem to get assembled but the bytes are not arranged
in the correct order (as defined by the instruction fields used).
Say I have an 16-bit unsigned immediate xy and a 8-bit signed
immediate n and my instruction format looks like this
0--7 8---------15 16-------23 24--31 32---------39
IIIIIIII yyyyyyyy xxxxxxxx IIIIIIII nnnnnnnn
I have defined instruction fields as
f-op1- 0,8 (0-7 bits)
f-op2 - 24,8 (24-31 bits)
f-uimm16 - 8,16 (8-23 bits)
f-simm8 - 32,8 (32-39 bits)
On seeing the disassembly, I find that both the opcodes are placed
contiguously (in 0-15 bits) and also the hi8 of 16-bit unsigned
immediate is filled with junk value.
Similar issue is seen when 16 bit immediate value is used in 24/32 bit
insns. Target is little-endian (in case that matters here).
Am I doing something wrong here?
> Presumably those IIIII's don't overlap - i.e., the hardware can tell
> from the first byte that it's a 2-byte instruction (and more opcode
> bits need to be fetched). In cgen, instruction opcodes need not be
> single fields nor contiguous; just specify one ifield per unique
> opcode piece.
Following example shows instruction formats with overlap
1-byte instruction:
IIIIIrrr - 5 bits opcode , 3 bits for register operand
2-byte instruction:
IIIIIsss IIIIIttt - 5-bit opcode, 3-bit register operand (source),
5-bit opcode, 3-bit register operand (destination)
I could not find a common bit pattern to identify the insn length. Can
this be handled using CGEN?
Thanks in advance.
Regards,
Usha Gupta